How to transcribe a video to text for free
Speech recognition on your own machine, with the timings stripped out at the end.
Transcription services are priced per minute because somebody has to pay for the GPUs doing the work, and they require an upload because the work happens on their machines. Both of those disappear when the model runs on your own computer — which it now can, in a browser tab, with quality that was research-grade a few years ago.
Open the subtitle editor → Transcribes on your own machine. Export plain text when you want the words without the timings.- 100% freeNo account, no minute quota, no paid tier.
- Nothing is uploadedYour video is read on your device and never reaches a server.
- No catch on the outputNo watermark, no length cap, no locked export formats.
- Updated in the openEvery fix is written down on the changelog page.
How to do it
Drop the video or audio file into the editor. It reads the audio track, downsamples it to what the speech model needs, and runs recognition locally — WebGPU if your browser has it, WebAssembly if not. Lines appear as it works, so you can start reading before it finishes.
When it is done, use Export text. That strips the timings and groups sentences into paragraphs, which is what you want for a transcript as opposed to a subtitle file. If you want the timings as well, export SRT instead and keep both.
Why local matters for this particular job
Think about what actually gets transcribed: research interviews, customer calls, therapy sessions, board meetings, medical consultations, legal depositions, internal all-hands recordings. This is close to a complete list of the material an organisation should not be uploading to a third-party service, and it is also close to a complete list of what transcription is used for.
Running locally removes the question entirely. There is no data processing agreement to read, no retention period to check, and no incident that can expose a file you never sent. For a lot of people this is not a preference, it is the only version of the job they are allowed to do.
What to expect from the output
For clear speech in a common language, expect most sentences to be right and a scatter of errors on names, acronyms and domain vocabulary. Punctuation is generally sensible. Multiple speakers are transcribed but not labelled — the model does not separate who is talking, so you add speaker markers yourself where it matters.
Difficult audio degrades the way you would expect: heavy accents, crosstalk, background music and phone-quality recordings all cost accuracy. The Accurate model setting is worth the extra time for any of those. Picking the language explicitly instead of leaving it on auto also helps noticeably on noisy recordings.
Long recordings are slow rather than impossible. An hour of audio is a coffee break on a machine with WebGPU and longer without it. There is no length limit and no charge either way.
Frequently asked questions
Is there a length limit?
No. The constraint is memory and patience: audio is reduced to 16 kHz mono, about 115 MB for an hour, and recognition time scales with length.
Can it tell speakers apart?
No. Speaker separation is a different model and is not implemented. The transcript is continuous text; you add speaker labels yourself.
How accurate is it?
Good on clear speech in common languages, weaker on names, jargon, accents and noisy recordings. Treat it as a strong first draft rather than a finished document.
Does it work with audio-only files?
Yes. MP3, M4A, WAV and similar all work — there is simply no picture to show alongside the text.
Why is it free when other services charge?
Because the computing happens on your hardware rather than someone's servers. There is no bandwidth or GPU bill to pass on. The site is paid for by ads on the page.